Detailed Analysis of Circuit-To-Hamiltonian Mappings Arxiv:1910.01481V1 [Quant-Ph] 2 Oct 2019

Detailed Analysis of Circuit-to-Hamiltonian Mappings

James D. Watson Department of Computer Science, University College London, UK

Abstract The circuit-to-Hamiltonian construction has found widespread use within the ﬁeld of Hamiltonian complexity, particularly for proving QMA-hardness results. In this work we examine the ground state energies of the Hamil- tonian for standard clock constructions and those which require dynamic initialisation. We put exponentially tight bounds on these ground state energies and also determine improved scaling bounds in the case where there is a constant probability of the computation being rejected. Furthermore, we prove a collection of results concerning the low-energy subspace of quantum walks on a line with energy penalties appearing at any point along the walk and introduce some general tools that may be useful for such analyses.

Contents

1 Introduction 2

2 Background and Previous Results 3 2.1 Preliminaries ...... 3 2.2 Feynman’s Construction ...... 4 2.3 Clock Constructions ...... 5 2.4 Spectra of Feynman-Kitaev Hamiltoninans and the Promise Gap . .6 2.5 Quantum Walks on a Line ...... 6 arXiv:1910.01481v1 [quant-ph] 2 Oct 2019 3 Main Results 7 3.1 Feynman-Kitaev Hamiltonians ...... 7 3.2 Quantum Walks on a Line ...... 8

1 4 Hamiltonian Analysis for Quantum Walks on a Line 9 4.1 Laplacian Matrix Analysis ...... 10 4.2 Endpoint Penalty Analysis ...... 13

5 Hamiltonian Analysis for Standard Form Hamiltonians 17 5.1 Standard Form Hamiltonians ...... 17 5.2 Standard-Form Hamiltonian Analysis ...... 18 5.3 Encoding (E)QMA Veriﬁcation in Standard Form Hamiltonians . . 24 5.3.1 EQMA Computations ...... 25 5.4 QMA Computation ...... 31

6 Eigenvalue Scaling with Constant Rejection Probability 35

7 Discussion and Outlook 39

8 Acknowledgements 40

1 Introduction

In the previous two decades there has been a union between condensed matter physics and complexity theory resulting in the new field of Hamiltonian complexity. The idea of Hamiltonian complexity is to study the properties of local Hamiltoni- ans from a complexity perspective, allowing us understand many-body quantum systems that may be to complicated to solve or are otherwise computationally in- tractable. Key to this field is the Feynman-Kitaev Hamiltonian construction which allows the evolution of a quantum circuit to be encoded in a many-body Hamil- tonian and was used in the seminal proof that the task of estimating the ground state of the local Hamiltonian problem is QMA-complete [KSV02]. This technique is often called a circuit-to-Hamiltonian mapping. Initial work used the mapping to prove QMA-hardness of the Local Hamiltonian problem for 5-local Hamiltoni- ans [KSV02] before it was then used to prove completeness of progressively more restricted classes of Hamiltonians. The culmination of the circuit to Hamiltonian mapping has been proving QMA-hardness of 1D Hamiltonians [Aha+09], and even 1D, translationally invariant, nearest neighbour Hamiltonians [GI09]. Perturbation gadgets developed first by [KKR04], combined with the circuit-to- Hamiltonian mapping, have been used to prove further sets of hardness results for systems on a lattice [OT08], and classify 2-local qubit Hamiltonians [CM13]. Work has also been done to classify the complexity of sampling [AA11], traversing the ground state subspace [GS15], counting low energy states [BFS11], excited states [JGL10], and finding the spectral gap [Amb14, GY16]. The circuit-to-Hamiltonian

2 construction has found usage in wide range of other areas in quantum information, including adiabatic quantum computation [Aha+08] and error correction [Boh+19]. Detailed analysis has been able to given improved bounds on the promise gap and eigenvalue scaling, not just for the standard constructions but also for non- uniform weights and for branching computations [CLN18, BC18, UHB17, BCO17]. However, the analysis of the gaps and eigenvalues are largely just scaling analysis. The aim of this paper is to pin down as closely as possible the eigenvalues and ground states of the Feynman-Kitaev construction. The contribution from this work is extending the analysis to constructions with more complex clock constructions which have been used in prior literature (in particular in the 1D case [CPW15, GI09]) which include clocks which require dynamic initialisation and do not have penalty terms only at the beginning and end. Fur- thermore, we pin down the ground state energy of these circuit-to-Hamiltonian mappings to within exponential precision. We also prove some potentially useful minimum eigenvalue bounds for Hamil- tonians of quantum walks on a line with penalties (or self-loops) which improve on bounds by [CLN18] and use diﬀerent methods. Using these, we prove tighter bounds for Hamiltonians which encode computations which reject with constant probability.

2 Background and Previous Results

In this section we give a brief overview of the Feynman-Kitaev circuit-to-Hamiltonian mapping and some previously known results about its properties.

2.1 Preliminaries

• if x ∈ AYES, ∃ |wi such that M accepts (x, |wi) with probability > 2/3.

• if x ∈ ANO then ∀ |wi, M accepts (x, |wi) with probability < 1/3. As an intermediate step we will ﬁnd it useful to prove results about EQMA (Exact-QMA) which is a zero-error quantum complexity class:

Deﬁnition 2.2 (EQMA). A promise problem A = (AYES,ANO) is in EQMA if there exist polynomials p and q and a QTM M such that for each instance x and

• if x ∈ AYES, ∃ |wi such that M accepts (x, |wi) with probability 1.

• if x ∈ ANO then ∀ |wi, M accepts (x, |wi) with probability 0 (i.e. always rejects).

We note that EQMA is not a particularly “natural” class and suffers the same ambiguities that EQP does (as defined in [BV97]). However, throughout the next few sections we will find it is easier prove results for the problem class EQMA as an intermediate step before using the EQMA results to prove results about QMA. We also take care to distinguish EQMA from the class NQP, defined in [ADH05], which has zero amplitude on the accept state if it is a rejecting instance, but is only required to have non-zero amplitude on the accept state when it is an accepting instance. It may still have non-zero amplitude on the reject state when the instance is an accepting instance whereas EQMA does not. Throughout the rest of the this work we will denote the matrix representing an N vertex path graph Laplacian as

 1 1  2 − 2 0 ...... 0 0 . − 1 1 − 1 .. 0   2 2   .   0 − 1 1 − 1 .   2 2   . .. 1 ..  (N)  . . − 1 .  ∆ =  2  . (2.1)  ......   . . . .   . . .   . .. .. 0     ...... 1   0 1 − 2  0 0 ...... 0 − 1 1 2 2 N×N

Furthermore, we denote an N × N matrix of zeros as 0N .

2.2 Feynman’s Construction

Consider a quantum circuit described by the unitary U = UN ...U2U1 (or alternatively a quantum TM evolving according to this set of unitaries). Further consider a set of qudits such that the total Hilbert space H = Hclock ⊗ Hregister. Hclock will contain a set of clock states |0i , |1i ,... |T − 1i, which will label the steps of the circuit after each unitary. Hregister will be the computational register state that is acted on by the unitaries.

4 We then design a Hamiltonian which has a “history state” ground state of the form

T −1 1 X |Ψi = √ |ti ⊗ (Ut ...U2U1) |φi , (2.2) T t=0 where |φi is the initial input state to the circuit. To do this we choose

H = Htrans + Hin, (2.3) where Htrans encodes the transitions/propagation of the circuit as

T −1 X Htrans = (|t + 1i ⊗ Ut − |ti)(ht + 1| ⊗ Ut − ht|), (2.4) t=0 and where the term Hin := |0i h0|clock ⊗|00 ... 0i h00 ... 0|ancillas applies a penalty if the ancilla qubits in the computational register do not start in the all zeros state. For a ﬁxed local Hilbert dimension, the clock states will generally be log(N)-local. However, typically there are ways of reducing the locality of the interaction to a constant [KSV02]. Often the circuit-to-Hamiltonian mapping is used with the aim of encoding a veriﬁcations circuit for a QMA problem. With this in mind one includes an output penalty

HFK := Hin + Hout + Htrans (2.5) where Hout = |T i hT | ⊗ Πout, where Πout is a projector onto a rejection ﬂag output by the quantum circuit (i.e. it penalises rejecting computations).

2.3 Clock Constructions The clock construction encoded in the Hamiltonian needs to be local and otherwise satisfy the constraints of the Hamiltonian. There are a multitude of clock constructions in the literature, notably including delocalised clocks [NT14] and translationally invariant clock constructions [GI09, CPW15]. Due to the constraints of encoding a clock construction into a translationally invariant Hamiltonian, these clocks require a dynamic initialisation and may undergo “bad” transitions (transitions to states that should not be allowed but cannot otherwise be excluded). We include them in our analysis below. Previous analyses of the circuit-to-Hamiltonian mapping have only achieved loose scaling bounds for such clock constructions, which we improve on here.

5 2.4 Spectra of Feynman-Kitaev Hamiltoninans and the Promise Gap The Feynman-Kitaev Hamiltonian is often invoked to prove QMA-hardness results. Here, we use the fact that if the Hamiltonian encodes a rejecting instance, it will have a high energy ground state, λ0 > β, otherwise if it encodes an accepting instance, it will have a low energy ground state λ0 < α, for β − α = O(1/ poly(n)) for a Hamiltonian on n qudits. This separation in α, β is known as the promise gap. In the original proof of QMA-hardness of the Local Hamiltonian problem, Ki- taev’s geometrical lemma is used to prove that the promise gap scales as Ω(T −3). Both [BC18] and [CLN18] improve on this to show that the standard Feynman- Kitaev Hamiltonian has a promise gap which scales a Ω(T −2). [BC18] also bounds the scaling for many non-uniform types of Hamiltonians which we will not be concerned with in this work. A further reason for interest in the promise gap comes from its relation to the quantum PCP conjecture [AAV13] – if it were possible to produce a suﬃciently large promise gap with a Feynman-Kitaev Hamiltonian then the PCP conjecture would follow. One can also consider alternative models of computation which branch, such as Quantum Thue Systems [BCO17]. It has been shown, using a generalisation of Kitaev’s geometrical lemma, that these models also have a promise gap Ω(N −3) where N is the number of vertices in the unitary labelled graph representing the computation (Lemma 44 of [BCO17]). Further work [BC18] shows that such constructions cannot be straightforwardly used to prove the quantum PCP conjecture. Finally, it is worth noting that all known history state constructions in the literature have been shown to have a spectral gap that closes as the length of the computation they encode increases [GC18]. A similar result was shown in [CB17]. In this work we pin down the promise gap for Feynman-Kitaev Hamiltonians with uniform weight transition rules to be within exponential precision of a ﬁxed function 1 − cos(π/2T ), where T is the runtime of the computation.

2.5 Quantum Walks on a Line It is well known that the Hamiltonian describing a particle hopping along a line of length T , where individual states are given by {|1i , |2i ,... |T i , }, is given by

T −1 X Hwalk = (|t + 1i − |ti)(ht + 1| − ht|) (2.6) t=1 = ∆(T ). (2.7)

6 This can be used to represent not just a particle propagating along a line, but a generic quantum process evolving. Throughout the rest of the paper, we will use a “walk” to refer to a graph Laplacian with weighted vertices. Our interest will be when the propagating process is a computation and the states represent the clock register labelling stages of the computation. In particular, the analysis of Feynman-Kitaev Hamiltonians, such as in equation 2.5, can often be mapped to quantum walks on a line. In the event that a computation gets an energy penalty it is often possible to show that the analysis becomes equivalent to analysing a Laplacian plus projectors for the relevant time steps as below X W †HwalkW = ∆(T ) + |ki hk| , (2.8) k∈K for some K ⊆ {1, 2,...,T }. With this in mind, [CLN18] put bounds on the scaling of the ground state energy for quantum walks on a line with penalties at their end points. In this work we improve on these bounds and introduce a set of techniques useful for analysing quantum walks on a line that have an energy penalty at a point which is not necessarily at the ends of the line.

3 Main Results

3.1 Feynman-Kitaev Hamiltonians We consider an extension of Feynman-Kitaev Hamiltonians to a more general class of Hamiltonians that have been considered before which we call Standard Form Hamiltonians that includes those which have “bad” clock transitions and clocks which may require a dynamic initialisation. Such clocks have appeared in [GI09], [CPW15], which are notable for encoding QTM rather than circuits. We then consider a QMA veriﬁcation computation encoded in the Hamiltonian and the associated minimum eigenvalues in both the accept and reject instances. We show that the ground state energy of a Hamiltonian encoding a veriﬁcation of a QMA YES or NO instance is given by the following theorem.

Theorem 3.1. The ground state energy of a standard-form Hamiltonian, HQMA ∈ B(Cd)⊗n, encoding the veriﬁcation computation of a QMA instance with total runtime T = poly(n) is bounded as

(YES) −O(poly(n)) 0 ≤λ0 HQMA ≤ e (3.1) π π 1 − cos − e−O(poly(n)) ≤λ H(NO) ≤ 1 − cos . (3.2) 2T 0 QMA 2T

7 Although these bounds do not give a better promise gap scaling compared to other known results — it remains Θ(T −2) as per [BC18] and [CLN18] — the above gives the minimum eigenvalue more precisely. Moreover, it gives bounds in the case where the clock has inherent bad transitions and so the Hamiltonian needs penalising terms and where the clock requires a dynamic initialisation — cases not covered by [CLN18] or [BC18]. Furthermore, we consider the case where the QMA acceptance probability is a constant and ampliﬁcation is no possible — for example for the class StoqMA.

(YES) Theorem 3.2 (YES Instance Upper Bound). Let HQMA encode the veriﬁcation of a YES QMA instance. Let η = O(1) be the maximum probability of rejection, then η 0 ≤ λ H(YES) = O . (3.3) 0 QMA T 2

This is an improvement on the bound found in [CLN18] of O(η/T ) and is expected to be tight.

3.2 Quantum Walks on a Line We present new eigenvalue bounds on Hamiltonians which encode quantum walks on a line where a penalty is applied, giving particular consideration to the case where the penalty is at the end of the walk. This analysis will later be applied to Hamiltonians encoding computation which is incorrect in some way. We also introduce the Uncoupling Lemma which allows the ground state energy of a quantum walk with a penalty on it to be analysed as two disjoint walks that share the penalty between them. We state and prove this in the next section. From this we are able to prove that if there is a walk on a line with a penalty term somewhere, then the lowest ground state energy is achieved with the penalty at one of the ends

Lemma 3.3. Consider the Hamiltonian

8 As an extension of this, we place bounds on the minimum energy eigenvalue for quantum walks with weight µ ≥ 0 penalties: H = ∆(T ) +µ |T i hT | and improve on the scaling in terms of µ: Theorem 3.4. For H = ∆(T ) + µ |T i hT | and µ = k/T , then for suﬃciently large T we have k λ (H) = Θ (3.6) 0 T 2 where k = O(1) is some constant.

4 Hamiltonian Analysis for Quantum Walks on a Line

Before we start further, we introduce a simple lemma that may ﬁnd use elsewhere. Given a quantum walk on a line that receives an energy penalty 1 on the kth step along its propagation, then the ground state energy can be bounded from below by decoupling the Hamiltonian into two disjoint quantum walk Hamiltonians and sharing the energy penalty between the two new walks. Lemma 4.1 (Uncoupling Lemma). Given a matrix H = ∆(T ) + |ki hk| , (4.1) then 1 1 H ≥ (∆(k−1) + |k − 1i hk − 1|) ⊕ (∆(T −k+1) + |ki hk|). (4.2) 4 2 Proof. We see that we can make a simple decomposition: H =∆(T ) + |ki hk| (4.3) 1 =(∆(k−1) + |k − 1i hk − 1|) ⊕ 0 + J + (0 ⊕ ∆(T −k+1)) (4.4) 4 T −k+1 k−1 where 1/4 −1/2 J = 0 ⊕ ⊕ 0 . (4.5) k−1 −1/2 1 T −k We note that J ≥ 0, hence 1 1 H ≥ (∆(k−1) + |k − 1i hk − 1|) ⊕ 0 + (0 ⊕ ∆(T −k+1) + |ki hk|). (4.6) 4 T −k+1 k−1 2

Using this we can directly bound eigenvalues of H.

9 4.1 Laplacian Matrix Analysis Before we begin we gather some results about tridiagonal toeplitz matrices. Let

b + γ c 0 ...... 0 α  .  ..   a b c 0   .   0 a b c .     . .. ..  (n)  . . a b .  A =   . (4.7)  ......   . . . .   . . .   . .. .. 0     . .   0 .. .. b c  β 0 ...... 0 a b + δ n×n then the following is true for speciﬁc instances of this family of matrices. Lemma 4.2 (Theorem 3.4 (iv) of [YC08]). Suppose that a = c 6= 0, γδ−αβ = −a2, and α + β = γ + δ = 0. Deﬁne (2k − 1)π θ = , k = 1, 2, . . . , n. (4.8) k 2n

(n) Then the eigenvalues of A are given by λk = b + 2a cos θk. Lemma 4.3 (Theorem 3.2 (viii) of [YC08]). Suppose that a = c 6= 0, αβ = γδ, α + β = 0 and γ + δ = a. Deﬁne (2k − 1)π θ = , k = 1, 2, . . . , n. (4.9) k 2n + 1

(n) Then the eigenvalues of A are given by λk = b + 2a cos θk. We note that ∆(n) is a special case of A(n) with a = c = −1/2, b = 1, γ = δ = −1/2 and α = β = 0. Hence the above lemmas will be useful in proving the following: Lemma 4.4 (Starting Penalty Lemma). Consider the T × T matrix

10 Proof. For the k = 1 case, Lemma 4.2 gives the minimum eigenvalue of the above as π λ(k=1) = 1 − cos . (4.12) 0 2T We now need to consider the k = 2, 3 cases separately. k = 2 : For the k = 2 we consider splitting the matrix up into two separate matrices in the following way:

 1/2 −1/2   1/4 −1/2  −1/2 2 −1/2    −1/2 1   −1/2 1      =   (4.13)  .   0   .. −1/2         −1/2 1 −1/2 −1/2 1/2 1/4 0   0 1 −1/2     −1/2 1    +  .  (4.14)  .. −1/2     −1/2 1 −1/2 −1/2 1/2 (4.15)

Note the ﬁrst matrix (i.e. the 2×2 block) is semi-positive deﬁnite (with eigenvalues λ ∈ {0, 5/4}), hence the following inequality holds:

 1/2 −1/2  1/4 0  −1/2 2 −1/2   0 1 −1/2       −1/2 1   −1/2 1       .  ≥  .  (4.16)  .. −1/2   .. −1/2       −1/2 1 −1/2  −1/2 1 −1/2 −1/2 1/2 −1/2 1/2

(T −1) 1 We then note that the bottom-right block of the matrix is ∆ + 2 |1i h1|. (k=2) π (k=1) From Lemma 4.3 this has a minimum eigenvalue λ0 = 1−cos 2T −1 > λ0 = 1 − cos π . Thus the k = 2 case has a larger minimum eigenvalue than the k = 1 case. 2T k = 3 : The k = 3 case follows similarly:

 1/2 −1/2   1/2 −1/2  −1/2 1 −1/2  −1/2 3/4 0       −1/2 2 −1/2   0 1 −1/2       −1/2 1  ≥  −1/2 1  (4.17)      ..   ..   . −1/2  . −1/2 −1/2 1/2 −1/2 1/2

11 where we have used the same method as√ in equation (4.13). The top-left block is positive deﬁnite with eigenvalues (5 ± 17)/8 > 0.1. From Lemma 4.3 the π (k=1) minimum eigenvalue of the right-hand side is then = 1 − cos 2T −3 > λ0 . Hence π λ(k=2) ≥ 1 − cos > λ(k=1). (4.18) 0 2T − 3 0

4 ≤ k ≤ bT/2c : We now consider the case for k ≥ 4. For this we will consider the matrix ∆(T ) + |ki hk| and split it into two block diagonal components:

 1/2 −1/2  0 0  −1/2 1  0 0       .   .   ..   ..   −1/2   0       −1/2 1 −1/2  =  0 1/4 −1/2       −1/2 2 −1/2   −1/2 1 0       −1/2 1 −1/2   0 0 0   .   .  −1/2 .. 0 .. (4.19)  1/4 −1/4   1/4 −1/4  −1/4 1/2  −1/4 1/2       .   .   ..   ..   −1/4   −1/4      +  −1/4 1/2 0  +  −1/4 1/4 0  .      0 1 −1/2   0 0 0       −1/2 1 −1/2   0 0 0   .   .  −1/2 .. 0 .. (4.20) Alternatively, we can write this as a block matrix decomposition:

1 (k−1) 1 1 (k−1) (T ) 2 ∆ + 4 |k − 1i hk − 1| 0 2 ∆ 0 ∆ + |ki hk| = (T −k+1) 1 + (4.21) 0 ∆ + 2 |ki hk| 0 0 0   1/4 −1/2  +   . (4.22)  −1/2 1  0

The second and third matrices are both positive semi-deﬁnite, and thus

1 (k−1) 1 (T ) 2 ∆ + 4 |k − 1i hk − 1| 0 ∆ + |ki hk| ≥ (T −k+1) 1 . (4.23) 0 ∆ + 2 |ki hk| Without loss of generality, we can now restrict to k ≤ T/2 (if k > T/2 we can do the same process as above, but swapping around the top-left and bottom-right blocks). We consider the two blocks separately:

12 Top-Left Block:

(j) 1 From Lemma 4.3 we ﬁnd that the minimum eigenvalue of ∆ + 2 |ji hj|, de- (j) noted λ0 , is π λ(j) = 1 − cos . (4.24) 0 2j + 1 Hence we know that the smallest possible minimum eigenvalue of the top-left block of right-hand side of 4.23 occurs when j = bT/2c. Thus 1 π λ(bT/2c) = 1 − cos (4.25) 0 2 2bT/2c + 1 π > 1 − cos = λ(k=1). (4.26) 2T 0

Bottom-Right Block:

Again, we ﬁnd the minimum eigenvalue of the bottom block is π λ = 1 − cos (4.27) 0 2(T − k + 1) + 1 π > 1 − cos (4.28) 2T (k=1) = λ0 , (4.29) where the ﬁrst to second line follows from the fact 4 ≤ k ≤ bT/2c.

(T ) Thus mink λ0 ∆ + |ki hk| is achieved for k = 1,T , for T > T0 = 3.

4.2 Endpoint Penalty Analysis In [CLN18] the spectra of quantum walks with end point penalties was considered. We do the same here and determine bounds for the ground state energies of these walks for a range of diﬀerent strength penalties. Our main object of study will be the Hamiltonian

(T ) HT (µ) = ∆ + µ |T i hT | , (4.30) where we will explore how the minimum eigenvalue varies as a function of T, µ.

Lemma 4.5. The eigenvalues of HT (µ) are the solutions of the equation r λ y (λ) − x (λ) µ T T = , (4.31) λ − 2 yT (λ) + xT (λ) 1 − µ

13 where

p T xT (λ) = (1 − λ − λ(λ − 2)) (4.32) p T yT (λ) = (1 − λ + λ(λ − 2)) . (4.33)

Proof. This follows from a standard recurrence relation for tridiagonal matrices: consider the characteristic equation det(H − λ1) = 0. We can use a standard continuant recurrence relation:

f0 = 1, (4.34)

f1 = 1/2 − λ, (4.35)

f2 = (1 − λ)f1 − (1/4)f0, (4.36) . . (4.37)

fn = (1/2 + µ − λ)fn−1 − (1/4)fn−1. (4.38)

Solving this gives the characteristic equation

T −1 −2 √ √ pT (λ) = √ (µ − 1)xT (λ) λ + µyT (λ) λ − 2 (4.39) λ − 2 = 0. (4.40)

Rearranging gives the formula as in the lemma statement.

Using the above lemma, we now prove properties of the eigenvalues of HT (µ):

Theorem 4.6. For µ = k/T , the minimum eigenvalue of HT (µ) is bounded by k k λ H = Θ , (4.41) 0 T T T 2 for k = O(1) and suﬃciently large T . Furthermore, for all µ and m ∈ N, 1 ≤ m ≤ T , (2m − 1)π λ (H (µ)) ≥ µ 1 − cos . (4.42) m T 2T

Proof. We ﬁrst take the characteristic equation and consider µ = 0, 1/2, 1 values. The eigenvalues corresponding to these values are known analytically by [YC08][Yue05]. Rearranging the characteristic equation gives r λ yT (λ) − xT (λ) µ gT (λ) = = , (4.43) λ − 2 yT (λ) + xT (λ) 1 − µ

14 which can be equivalently written as

p(µ=0)(λ) µ T = . (4.44) (µ=1) 1 − µ pT (λ)

A sketch of gT (λ) can be seen in ﬁgure 1 for T = 7. The eigenvalues for µ = 0, 1 are known known analytically, hence it is known (µ=1) (2k−1)π (µ=0) pT (λ) has zeros at λk = 1 − cos 2T for k = 1, 2,...,T , and pT (λ) has 2kπ zeros at λk = 1−cos (2T +1) , k = 0, 1,...,T −1. From this, we ﬁnd that gT (λ) has (2k−1)π 2kπ poles at λk = 1−cos 2T for k = 1, 2,...,T , and zeros at λk = 1−cos (2T +1) , k = 0, 1,...,T − 1 which are the eigenvalues of the µ = 0 case. Furthermore, we know the eigenvalues of the µ = 1/2 case and know that this occurs when the left-hand side of equation 4.43 is equal to one. The eigenvalues (2k−1)π for this case are γk = 1 − cos 2T +1 , k = 1, 2,...,T , hence g(γk) = 1 [Yue05]. Hence we know gT (λ) must look like Fig. 1. By examining this, we can trivially put an upper bound on many of the eigenvalues, including the smallest, as O(T −2).

0 μ /( 1 - μ )

-2

-4

0.0 0.5 1.0 1.5 2.0 λ

Figure 1: The red line represents function g7(λ). The blue horizontal line corre- sponds to µ = 1/2.

15 To ﬁnd solutions we consider where gT intersects horizontal lines of µ/(1 − µ) (see Fig. 1 for an example). In particular we are interested in the case of large T when µ is small. We consider the cases where µ = k/T for some k = O(1). We consider the expansion of gT around T = ∞ in powers of 1/T to get r k k √ 1 g = tan 2k + O(T −2). (4.45) T T 2 2 T

We want to determine when this is equal to µ/(1 − µ) = k/(T − k). Hence for suﬃciently large T and k = O(1), we can write

k k k c < g < c0 (4.46) T T T 2 T for some c, c0 = O(1), c0 > c. This gives

k k λ H = Θ . (4.47) 0 T T T 2

Lower Bound Finally we lower bound the eigenvalues:

∆(T ) + µ |T i hT | = µ(∆(T ) + |T i hT |) − (1 − µ)∆(T ) (4.48) ≥ µ(∆(T ) + |T i hT |) (4.49)

Hence, using the above inequality, we have that

(T ) (T ) λm(∆ + µ |T i hT |) ≥ µλm(∆ + |T i hT |) (4.50) (2m − 1)π ≥ µ 1 − cos . (4.51) 2T

Although we do not prove it, numerical analysis suggests that the gradient of 3 gµ(λ) at µ = 1/2 scales as O(T ). This suggests that for general 0 < µ < 1/2 and T −3 2 large T , the best bound achievable is λ0(HT (µ)) = O(µ /T ). We note the ﬁrst bound in the lemma statement is an extension on [CLN18] where it was shown that λ0(HT (µ)) = O(µ/T ). While this would give us the same upper bound as our result, it does not give the same lower bound. We note that the lower Ω(k/T 2) result here is only true for µ = k/T, k = O(1) and does not apply for the case µ = O(1) in general.

16 5 Hamiltonian Analysis for Standard Form Hamil- tonians

In this section we consider Hamiltonians which encode computation with uniform weight transition rules. We also restrict ourselves to computations which do not branch — given a basis state there is at most one transition rule that applies to it. We further restrict ourselves to the analysis of computations which do not branch into multiple tracks (such as those described by unitary labelled graphs as per [BCO17]). We ﬁrst consider computations which have a deterministically accepted or rejected output: Exact-QMA (EQMA), as deﬁne in Def. 2.2. This will give us the relevant tools for examining Hamiltonians that encode QMA instances.

5.1 Standard Form Hamiltonians The Hamiltonians we are interested in will fit into specific class of Hamiltonians which we call “standard form Hamiltonians”. The idea will be that we encode the verification of problem instances in these Hamiltonians in a history state construction, as per [KSV02]. Thus, as usual for these constructions, this class of Hamiltonians will contain three types of terms. The first is transition terms which force the evolution from one state to the next. The second type of terms are penalty terms which act to assign an energy penalty to any states which are not allowed by the computation. The final set of terms are computational penalty terms which penalise computations that are not correctly initialised or result in a NO instance after the computation has been run. We label this class ‘Standard-Form Hamil- tonians’, the definition of which is a modification to the class of the same name from [CPW15].

Deﬁnition 5.1 (Standard Basis States, from Section 4.1 of [CPW15]). Let the single site Hilbert space be H = ⊗iHi and ﬁx some orthonormal basis for the single site Hilbert space. Then a Standard Basis State for H⊗L are product states over the single site basis.

We now deﬁne standard-form Hamiltonians — extending the deﬁnition from [CPW15]:

Deﬁnition 5.2 (Standard-form Hamiltonian, deﬁnition extended from [CPW15]). We say that a Hamiltonian H = Htrans + Hpen + Hin + Hout acting on a Hilbert C Q ⊗L C ⊗L Q ⊗L space H = (C ⊗ C ) = (C ) ⊗ (C ) =: HC ⊗ HQ is of standard form PL−1 (i,i+1) if Htrans,pen,in,out = i=1 htrans,pen,in,out, and the local interactions htrans,pen,in,out satisfy the following conditions:

17 C Q ⊗2 1. htrans ∈ B (C ⊗ C ) is a sum of transition rule terms, where all the transition rules act diagonally on CC ⊗ CC in the following sense. Given standard basis states a, b, c, d ∈ CC , exactly one of the following holds:

• there is no transition from ab to cd at all; or C Q Q • a, b, c, d ∈ C and there exists a unitary Uabcd acting on C ⊗C together i CQ CQ with an orthonormal basis {|ψabcdi}i for ⊗ , both depending only on a, b, c, d, such that the transition rules from ab to cd appearing in i i htrans are exactly |abi |ψabcdi → |cdi Uabcd |ψabcdi for all i. There is then a corresponding term in the Hamiltonian of the form (|cdi ⊗ Uabcd − † |abi)(hcd| ⊗ Uabcd − hab|).

C Q ⊗2 2. hpen ∈ B (C ⊗ C ) is a sum of penalty terms which act non-trivially C ⊗2 only on (C ) and are diagonal in the standard basis, such that hpen = P 1 (ab) Illegal |abiC hab| ⊗ Q, where (ab) are members of a disallowed/illegal subspace.

P CC ⊗2 3. hin = ab |abi hab|C ⊗ Πab, where |abi hab|C ∈ ( ) is a projector onto CC ⊗2 (in) CQ ⊗2 CQ ⊗2 ( ) basis states, and Πab ∈ ( ) are orthogonal projectors onto ( ) basis states.

CC ⊗2 CC ⊗2 4. hout = |xyi hxy|C ⊗Πxy, where |xyi hxy|C ∈ ( ) is a projector onto ( ) (in) Q ⊗2 Q ⊗2 basis states, and Πxy ∈ (C ) are orthogonal projectors onto (C ) basis states.

We note that although hout and hin have essentially the same form, they will play a diﬀerent role later in the proof. We reserve hout for penalties applied to the output of the QTM only.

5.2 Standard-Form Hamiltonian Analysis In this section we exactly determine the minimum eigenvalues of a standard-form Hamiltonian which has a ground state that encodes the veriﬁcation of either a YES or NO EQMA problem instance. This is then used to bound the minimum eigenvalues for Hamiltonians encoding QMA computations. We begin with several deﬁnitions:

Deﬁnition 5.3 (Legal and Illegal Pairs and States, from [CPW15]). The pair ab 1 is an illegal pair if the penalty term |abi hab|C ⊗ Q is in the support of the Hpen component of the Hamiltonian. If a pair is not illegal, it is legal. We call a standard basis state legal if it does not contain any illegal pairs, and illegal otherwise.

18 Then the following is a straightforward extension of Lemma 42 of [CPW15] with Hin and Hout terms included. Lemma 5.4 (Invariant subspaces, extended from Lemma 42 of [CPW15]). Let Htrans, Hpen, Hin and Hout deﬁne a standard-form Hamiltonian as deﬁned in Def. 5.2. Let S = {Si} be a partition of the standard basis states of HC into minimal subsets Si that are closed under the transition rules (where a transition rule C ⊗2 |abi |ψi → |cdi Uabcd |ψi acts on HC by restriction to (C ) , i.e. it acts CD CD L as ab → cd). Then H = ( S KSi ) ⊗ HQ decomposes into invariant subspaces

KSi ⊗ HQ of H = Hpen + Htrans + Hin + Hout where KSi is spanned by Si.

This is useful as it allows us to divide up the Hilbert space of Htrans + Hpen + Hin + Hout into invariant subspaces in which each state has at most one transition applied to it in the forwards and backwards directions. We can then study the minimum eigenvalues of these subspaces separately. We use a modiﬁed version of the Clairvoyance Lemma from [CPW15] to do this. However, before we can prove this, we need to introduce the following lemma that allows us to put a projector in a stoquastic form:

Lemma 5.5 (Lemma 20 of [BC18]). Let M ∈ Cd×d be a projector. Then for any 1 ≤ s < d, there exists a block-diagonal unitary V = V 0 ⊕ V 00 with dim V 0 = s such that D = V †MV is stoquastic. Furthermore, V can be chosen such that, if we denote the rank of the upper-left s × s block with ra and the complementary lower-right block rank with rb, then Dij 6= 0 if and only if

1. i = j and j ≤ ra

2. or s ≤ i = j ≤ s + rb

3. or s ≤ j = s + i ≤ min{ra, rb}

4. or s ≤ i = s + j ≤ min{ra, rb}.

More intuitively, we can write

D D V †MV = D = aa ab , (5.1) Dba Dbb where Daa and Dbb are real, diagonal matrices with rank ra and rb respectively, and dimensions s × s and (|HQ| − s) × (|HQ| − s) respectively. Furthermore, Dab and Dba are real, negative, diagonal matrices with rank Dab = rank Dba = min{ra, rb}. Its form can be seen explicitly as the red part in ﬁgure 2. Our statement of the modiﬁed version of the Clairvoyance Lemma is

19 Lemma 5.6 (Clairvoyance Lemma, extended from Lemma 43 of [CPW15]). Let H = Htrans + Hpen + Hin + Hout be a standard-form Hamiltonian, as deﬁned in Def. 5.2, and let KS be deﬁned as in Lemma 5.4. Let λ0(KS) denote the minimum eigenvalue of the restriction H|KS ⊗HQ of H = Htrans + Hpen + Hin + Hout to the invariant subspace KS ⊗ HQ. Assume that there exists a subset W of standard basis states for HC with the following properties:

1. All legal standard basis states for HC are contained in W. 2. W is closed with respect to the transition rules.

3. At most one transition rule applies in each direction to any state in W. Furthermore, there exists an ordering on the states in each S such that the forwards transition (if it exists) is from |ti → |t + 1i and the backwards transition (if it exists) is |ti → |t − 1i.

4. For any subset S ⊆ W that contains only legal states, there exists at least one state to which no backwards transition applies and one state to which no forwards transition applies. Furthermore, the unitaries associated with each transition |ti → |t + 1i are Ut = 1Q, for 0 ≤ t ≤ Tinit − 1. Also, both the ﬁnal state |T i, and whether a state |ti has t ≤ Tinit, is detectable by a 2-local, translationally invariant projector acting only on nearest neighbour qudits.

Then each subspace KS falls into one of the following categories: 1 1. S contains only illegal states, and H|KS ⊗HQ ≥ . 2. S contains both legal and illegal states, and

† M (|S|) X W H|KS ⊗HQ W ≥ ∆ + |ki hk| (5.2)

i |ki∈Ki

where P |ki hk| := H | and K is some non-empty set of basis |ki∈Ki pen KS ⊗HQ i states and W is some unitary.

3. S contains only legal states, then there exists a unitary R = W (1C ⊗ (X ⊕

Y )Q) that puts H|KS ⊗HQ in the form † Haa Hab R H|KS ⊗HQ R = † , (5.3) Hab Hbb PTinit−1 (in) where, deﬁning G := supp t=0 Πt and s := dim G,

20 • X : G → G. • Y : Gc → Gc.

• Haa is an s × s matrix.

• Haa,Hbb ≥ 0 and are rank ra, rb respectively.

• Haa has the form

Tinit−1 M (|S|) X † Haa = ∆ + αi ||S| − 1i h|S| − 1| + |ti ht| ⊗ X Πt|GX. i t=0 (5.4)

• Hbb is a tridiagonal, stoquastic matrix of the form

M (|S|) Hbb = (∆ + βi ||S| − 1i h|S| − 1|). (5.5) i

• Hab = Hba is a real, negative diagonal matrix with rank min{ra, rb}. M Hab = Hba = γi ||S| − 1i h|S| − 1| . (5.6) i where either we get pairings between the blocks such that p αi γi 1 − µi − µi(1 − µi) 1 0 = p or , (5.7) γi βi − µi(1 − µi) µi 0 1

for 0 ≤ µi ≤ 1, or we get unpaired values of αi = 0, 1 or βi = 0, 1 for which we have no associated value of γi.

Proof. The case of type 1 subspaces is straightforward as hx|C hψ|Q Hpen |xiC |ψiQ ≥ 1 1 for any illegal standard basis state |xiC . Thus, H|KS ⊗HQ ≥ . We consider subspaces of type 2 and 3. To begin with, we initially follow the analysis from [CPW15]. Consider the directed graph of states in W formed by adding a directed edge between pairs of states connected by transition rules. By assumption, only one transition rule applies in each direction to any state in W, so the graph consists of a union of disjoint paths (which could be loops in case 2). Minimality of S (Lemma 5.4) implies that S consists of a single such connected path. Let t = 0,..., |S| − 1 denote the states in S enumerated in the order induced by the directed graph. Htrans then acts on the subspace KS ⊗ HQ as

T −1 X 1 H | = |ti ht| ⊗ 1 + |t + 1i ht + 1| ⊗ 1 − |t + 1iht| ⊗ U − |tiht + 1| ⊗ U † trans KS ⊗HQ 2 t t t=0 (5.8)

21 where T = |S| − 1 if the path in S is a loop, otherwise T = |S| − 2. Whether we have a loop or not, we have

|S|−2 X 1 H | ≥ |ti ht| ⊗ 1 + |t + 1i ht + 1| ⊗ 1 − |t + 1iht| ⊗ U − |tiht + 1| ⊗ U † trans KS ⊗HQ 2 t t t=0 (5.9)

:= Hpath (5.10)

Equality arises when the path is not a loop. Furthermore, the ordering of the states in S means we can write

T −1 initX Hin = |ti ht| ⊗ Πt (5.11) t=0

Hout = |T i hT | ⊗ Πout. (5.12)

Now deﬁne

|S|−2 t X Y † 1 W := |ti ht| ⊗ Ui + ||S| − 1i h|S| − 1| ⊗ Q. (5.13) t=0 i=0

QT −1 where we have deﬁned U := j=0 Uj and have written out Hin,Hout explicitly. † Again, equality holds when the path is not a loop. We see that W Hpen|KS ⊗HQ W =

Hpen|KS ⊗HQ as, by Def. 5.2, Hpen only acts non-trivially on HC while the unitaries Ut act non-trivially only on HQ. Additionally, S is deﬁned to be minimal. We now consider subspaces of type 2 and 3 separately.

Type 2 Subspaces By deﬁnition computations in type 2 subspaces must evolve to an illegal state at some point and hence Hpen|KS ⊗HQ must have some support on any subspace of this type. Noting that the last two terms in expression 5.14 are positive semi-deﬁnite

22 hence we can remove them to give the following,

† (|S|) 1 W H|KS ⊗HQ W ≥ ∆ ⊗ Q + Hpen|KS ⊗HQ . (5.15) † M (|S|) X W H|KS ⊗HQ W ≥ (∆ + |ki hk|) (5.16)

i |ki∈Ki for some non-empty set of basis states Ki.

Type 3 Subspaces

By deﬁnition, all the states in type 3 subspaces are legal, and hence Hpen|KS ⊗HQ = 0 in this subspace. Furthermore, the states in S cannot form a loop by condition 4 in the lemma statement. Thus the Hamiltonian takes the form

We then use Lemma 5.5 to deﬁne a unitary (1C ⊗V ), where V = X ⊕Y , which † † puts V U ΠtUV into the form described in Lemma 5.5. We choose X to be an PTinit−1 s × s unitary with the same support as t=0 Πt. This gives

PTinit−1 and Maa is an s × s matrix with the same support as t=0 Πt. We can then † decompose (UV ) ΠoutUV as a series of block diagonal terms

† M (UV ) ΠoutUV = Pi, (5.21) i where either λi −|ξi| Pi = (5.22) −|ξi| µi or Pi is equal to the 1 × 1 matrices Pi = 1 or Pi = 0. Each non-zero Pi must be a projector since Πout is a projector and is block diagonal in the Pi. The resulting

23 Figure 2: The conjugated Hamiltonian in restricted to a type 3 subspace. The conjugated Hout terms — in red — couple blocks in the support of Hin to a block outside the support. The conjugated Hin terms — represented by the term on (T ) the right — couple ∆ blocks in the Πin subspace together, where the coupling occurs on the ﬁrst Tinit terms.

conjugated Hamiltonian can be seen in ﬁgure 2. If Pi is rank 2, then Pi = 12 (as this is the only rank 2, 2 × 2 projector). If it is a rank 1, 2 × 2 matrix then we can parametrise it in terms of a single value: p 1 − µi − µi(1 − µi) Pi = p (5.23) − µi(1 − µi) µi (we see this from the fact P is a rank 1 projector and hence we must be able to i √ √ write it as Pi = |χi hχ| for |χi = 1 − µi |0i− µi |1i). This is the form as claimed in the lemma statement.

5.3 Encoding (E)QMA Veriﬁcation in Standard Form Hamil- tonians The Clairvoyance Lemma proves properties for a general standard form Hamilto- ⊗n nians under certain assumptions. We identify the basis states in HC as “clock states”, and as per Def. 5.2, the transitions between these states are deterministic. The clock states, together with the transition rules between them, form the clock. All penalties acting on the clocks states are diagonal. There is a least one set of clock states for which there is an evolution which does not contain any illegal states and has a start and ﬁnish state: we called this a valid evolution. Naturally we label the clock states in order as |0i → |1i → · · · → |ti → · · · → |T i.

24 To encode a Quantum Turing Machine in a Hamiltonian, we choose Ut to correspond to a transition from the transition rule table, and a design a clock construction with an associated classical Hilbert space which labels the Quantum Turing Machine operations. Although we will not explicitly state a clock construction, we note that constructions similar to the unary clock in [GI09] and the base ξ clock in [CPW15]both satisfy the assumptions in the Clairvoyance Lemma. In both these cases, if H ∈ B(Cd)⊗L, then the computations these Hamiltonians encode have runtimes O(L2) and O(LξL log(L)) respectively, for some ξ we are free to choose. These clocks have a dynamic initialisation and hence much of the previous analysis in [BC18] and [CLN18] does not apply to them. We further introduce a set of other properties that these clocks share that we will ﬁnd useful. We call this class of clocks “Standard Form Clocks”. Assumption 5.7 (Standard Form Clock Properties). We assume the standard form clock construction for a standard form Hamiltonian H ∈ B(Cd)⊗n has the following properties: • Satisﬁes assumptions 1-4 in the Clairvoyance Lemma (Lemma 5.6). • The total runtime of the clock is T = T (n) = Ω(n2). • For any set of basis states S which contains at least one illegal state, then given any legal state k ∈ S, k will evolve to an illegal state within O(n) transitions.

1/2 • The initialisation time is bounded as Tinit = O(n) = O(T ).

5.3.1 EQMA Computations We now consider a standard form Hamiltonian which will encode the time evolution of an EQMA verifier computation. Lemma 5.8 (EQMA Clairvoyance Lemma, extended from Lemma 43 of [CPW15]). Let HEQMA = Htrans + Hpen + Hin + Hout be a Standard Form Hamiltonian encoding the verification of a EQMA problem instance with a standard form clock. Let the subspaces KS be defined as in Lemma 5.4. Let λ0(KS) denote the minimum eigenvalue of the restriction H|KS ⊗HQ of HEQMA = Htrans + Hpen + Hin + Hout to the invariant subspace KS ⊗ HQ. Then each KS falls into one of the following categories (corresponding to the same categories in Lemma 5.6):

1. S contains only illegal states, and λ0(KS) ≥ 1.

π 2. S contains both legal and illegal states, and λ0(KS) ≥ 1 − cos O(n) .

25 3. S contains only legal states, and if the Hamiltonian’s ground state encodes the veriﬁcation of YES or NO instance, then ( 0 Y ES instance λ (K ) = 0 S π 1 − cos 2T NO instance, where T (n) = Ω(n2) is the runtime of the standard form clock construction. Furthermore, the Hamiltonian restricted to this subspace takes the form given in Lemma 5.6 for subspaces of type 3, where Hab = Hba = 0, and

M (T ) Hbb = (∆ + δi) (5.24) i

where δi = 1 always for NO instances. For YES instances δi = 0 for at least one case, and is 1 otherwise. Proof. By assumption, the standard form clock satisfies all the assumptions for the Clairvoyance Lemma (Lemma 5.6) to hold, hence we can apply it straightforwardly. We now consider the three different types of subspaces as defined in the Clair- voyance Lemma. The case of subspace 1 follows straightforwardly from the result about subspace 1 in Lemma 5.6. a

Type 3 Subspaces:

Here we consider subspaces which contain only legal states (as per point 3 point 3 of the lemma statement). From the analysis in Lemma 5.6 that the Hamiltonian takes the form

EQMA NO Instances

We now consider the minimum eigenvalues in the case that the ground state encodes the veriﬁcation of an NO EQMA instance. We will need the following lemma from [BC18]:

26 Lemma 5.9 (Lemma 15 of [BC18]). Let Ut be the unitary representing the evolu- th QT tion of the computation at its t step. Let U = t=0 Ut encode the veriﬁcation of † PTinit a NO instance. Then Dbb, deﬁned as R ΠoutR restricted to ker t=1 Πt, has full rank. For a NO EQMA instance, the probability of any input being rejected is 1. We now realise that for a NO instance the probability of the circuit rejecting a PTinit=1 Cd ⊗L † correctly initialised input |xi ∈ ker( t=0 Πt) ⊂ ( ) is hx| U ΠoutU |xi = 0 0 0 hx | Dbb |x i = 1 for an EQMA instance. If we choose |x i ∈ ker Hin then this, in combination with Dbb having maximum rank, implies µi = 1, ∀i. L If we rearrange the rows and columns and write D = i Pi, then µi = 1 implies either Pi|Gc = 1 or for 2 × 2 matrices Pi = 12 or

0 0 P = . (5.27) i 0 1

The part of the conjugated Hamiltonian represented by Hbb in the Clairvoyance Lemma (Lemma 5.6) now decouples into a set of T × T blocks (see figure 2). We † also note Hab = Hba = 0. We now know that the lowest eigenvalue of Hbb must belong to one of these separate T × T blocks, or lie in the upper-left block Haa. We now consider these two possibilities. These T × T blocks are penalised by Hout only; let Kout represent one of these blocks, then supp(Kout) ∩ supp(Hin) = ∅ (where G is defined in Lemma 5.6). As a result they must take the form   (T ) (T ) Kout = ∆ + |T i hT | =  ∆  . (5.28) 1 PTinit−1 We now want to show that the subspace supp t=0 Πt (i.e. the subspace penalised by Hin) must have a minimum eigenvalue greater than or equal to the minimum eigenvalue of K . To consider the blocks penalised by H we first label out in PTinit−1 supp t=0 Πt =: G. We note D|G = (1 − µ1) |1i h1| + (1 − µ2) |2i h2| + ··· = L i(1 − µi), where µi = 0 or 1. Then the Hamiltonian restricted to this subspace is

27 Tinit−1 † (T ) † X M R HR|G = ∆ ⊗ 1G + (R |ti ht| ⊗ ΠtR)|G + |T i hT | ⊗ (1 − µi) t=0 i (5.29)

Tinit−1 (T ) † X ≥ ∆ ⊗ 1G + (R |ti ht| ⊗ ΠtR)|G (5.30) t=0

We then recall that R = W (1C ⊗ (X ⊕ Y )Q), hence we can conjugate with the inverse 1C ⊗ XQ = R|G to get

Tinit−1 † (T ) X W HW |G ≥ ∆ ⊗ 1G + |ti ht| ⊗ Πt|G (5.31) t=0 s M X ≥ ∆(T ) + |ki hk| . (5.32)

i=1 |ki∈Zi where Zi is a non-empty set of basis elements corresponding to the elements penalised by Hin, for 0 ≤ k ≤ Tinit − 1. Going from equation (5.31) to (5.32) we have used the fact that the Hamiltonian decomposes into blocks: one block in G c L and the other in G . Then the matrix i(1 − µi) is positive semi-deﬁnite. Note that the matrix in equation (5.32) is a block diagonal matrix with blocks of the form (T ) X Kin(Zi) = ∆ + |ki hk| (5.33)

|ki∈Zi

We now want to show that minimum eigenvalue of blocks of the form Kin(Zi) is larger than those of Kout ∀k. To do this we use Lemma 4.4 derived earlier (T ) First realise that Kin(Zi) ≥ ∆ + |ji hj|, where j is the smallest integer such that |ji ∈ Zi. To see that a Kout block always exists we note that it must be possible to choose a state |xi ∈ ker Hin for a non-trivial kernel, and for a NO EQMA instance this must correspond to a Kout block. From Lemma 4.4 we see that the minimum eigenvalue of ∆(T ) + |ki hk| occurs for k = 1, which is equal to (T ) the minimum eigenvalue of Kout blocks. Hence Kin(Zi) ≥ ∆ + |ji hj| ≥ Kout. (2m−1)π From Lemma 4.2 we ﬁnd that the eigenvalues of Kout blocks are 1−cos 2T , for m = 1, 2..., T , thus giving a minimum eigenvalue of a NO EQMA instance as: π λ (K ) = 1 − cos (5.34) 0 S 2T

28 EQMA YES Instances PTinit−1 For an EQMA YES instance we assume that ker t=0 Πt is non-trivial. Then, by deﬁnition, for a YES EQMA instance, there exists a state that is in PTinit † ker t=0 Πt + U ΠoutU . Thus there exists a block with µi = 0, hence only contains ∆(T ). By standard analysis we see that the corresponding minimum eigenvalue eigenspace for YES instances is

Type 2 Subspaces:

We now consider subspaces which contain both legal and illegal states. From the Clairvoyance Lemma (Lemma 5.6) we have that the Hamiltonian after conjugation takes the form: † M (|S|) X W H|KS ⊗HQ W ≥ ∆ + |ki hk| (5.36)

i |ki∈Zi where P |ki hk| := H | . We note that here each basis state within a |ki∈Zi pen KS ⊗HQ KS ⊗HQ block represents a diﬀerent time step as the computation propagates. We want to lower-bound the energy of these KS ⊗ HQ subspaces such that they have energy larger than subspaces of type 3. Thus we consider the clock properties assumption (Assumption 5.7) which tells us that any state in subspace 2 must reach an illegal state in O(n) steps forwards or backwards. Thus for |ki ∈ Zi there must be another state |ri ∈ Zi, such that r ≤ k + O(n) (unless |S| < k + O(n)) and similarly in the backwards direction. Now consider the right-hand side of eq. (5.36). This can be decomposed into sets of shorter 1D walks of length `i ≤ O(n) such that each of these shorter paths

29 contains an penalty term in its ﬁrst element, thus allowing us to write (|S|) X M (`j ) X ∆ + |ki hk| = ∆ + |mji hmj| + Ej (5.37)

|ki∈Zi j M (`j ) ≥ ∆ + |mji hmj| (5.38) j

(`i) where |mji ∈ Zi is some basis state corresponding to the ﬁrst element of ∆ , and 1/2 −1/2 E = . (5.39) j −1/2 1/2

The inequality comes from the fact the terms Ej are are positive semi-definite. Since ì ≤ O(n), and using Lemma 4.4, we can bound the minimum eigenvalue of each of these matrices as π λ (K ) ≥ 1 − cos . (5.40) 0 S O(n) For sufficiently large n, this is always larger than the minimum eigenvalues of type-3 subspaces. So far we have found the minimum eigenvalue of the standard form Hamiltonian encoding the verification of an EQMA instances. We now consider the form of the ground states themselves. Lemma 5.10 (EQMA Ground States). Let H ∈ B(C⊗n) be a standard form Hamiltonian as described in Def. 5.2. Let the Hamiltonian encode the verification Qt of an EQMA instance and define |ψti = j=0 Ut |ψ0i, for some initial state |ψ0i. Then the ground states for the YES and NO instances take the form

T −1 1 X |νYESi = √ |ti |ψti (5.41) T t=0 T −1 X (t + 1)π πt |ν i = sin − sin |ti |ψ i , (5.42) NO 2T 2T t t=0 T −1 X (2t + 1)π πt = 2 cos sin |ti |ψ i . (5.43) 2T 2T t t=0 Proof. We see that the ground state energies for these two cases correspond to Hamiltonians of the form of T × T matrices

(YES) (T ) HEQMA|KS ⊗HQ = ∆ , (5.44)

30 which is well known to have a uniform superposition as its ground state. For the NO instance, we see

(NO) (T ) HEQMA|KS ⊗HQ = ∆ + |T i hT | . (5.45) The ground state of this is given in [YC08].

5.4 QMA Computation We now consider a Standard-Form Hamiltonian encoding the veriﬁcation of a QMA problem instance and its associated eigenvalues. Before we do so we introduce the following lemma:

Lemma 5.11 (Lemma 4 of [KKR04]). Let H1 and H2 be two Hamiltonians with eigenvalues µ1 ≤ µ2 ≤ ... and σ1 ≤ σ2 ≤ .... Then, for all j, |µj −σj| ≤ ||H1 −H2||.

Lemma 5.12 (QMA Clairvoyance Lemma). Let H = Htrans + Hpen + Hin + Hout be a Standard-Form Hamiltonian encoding the evolution of a QMA veriﬁer. Let the subspaces KS be deﬁned as in Lemma 5.4 and let λ0(KS) denote the minimum eigenvalue of the restriction H|KS ⊗HQ of H = Htrans + Hpen + Hin + Hout to the invariant subspace KS ⊗ HQ. Then each KS falls into one of the following categories (corresponding to the same categories in Lemma 5.6):

1. S contains only illegal states, and λ0(KS) ≥ 1.

π 2. S contains both legal and illegal states., and λ0(KS) ≥ 1 − cos 8L . 3. S contains only legal states. Deﬁne η to be the probability of error for QMA, as per Def. 2.1. Let (Y ES/NO) HQMA represents a Hamiltonian encoding the veriﬁcation of a YES/NO instance, then its minimum eigenvalue is bounded by

(YES) 1/2 0 ≤λ0 HQMA |KS ⊗HQ ≤ η (5.46) π π 1 − cos − η1/2 ≤λ H(NO)| ≤ 1 − cos (5.47) 2T 0 QMA KS ⊗HQ 2T

Proof. The Hamiltonian is of standard form and hence satisﬁes the assumptions of the Clairvoyance Lemma (Lemma 5.6). The statements for subspaces of types 1 and 2 follow directly from Lemma 5.6 by the same reasoning as the EQMA case in Lemma 5.8. We now consider subspaces of type 3. To do this we deﬁne an “EQMA Hamil- ˜ tonian” HEQMA that has the same eigenvalues as HQMA would have if all its

31 verification computations were computed deterministically. We then bound the eigenvalues of type 3 subspaces relative to this EQMA Hamiltonian. a Defining the EQMA Hamiltonian First we consider a standard form Hamiltonian which encodes the verification a QMA instance, HQMA. From it, we define an EQMA Hamiltonian which cor- responds to it. Following from the Clairvoyance Lemma (Lemma 5.6), we can 1 conjugate HQMA|KS ⊗HQ by the unitary R = W ( C ⊗ (X ⊕ Y )Q) to put it in the following form:

† (T ) 1 † R HQMA|KS ⊗HQ R =∆ ⊗ Q + R Hin|KS ⊗HQ R (5.48) + |T i hT | ⊗ DQMA, (5.49)

QMA L QMA QMA where D = Pi , where Pi are the matrices described in the Clair- voyance (Lemma 5.6). We then deﬁne the corresponding EQMA Hamiltonian to be:

EQMA QMA QMA where Pi is of the same dimension as the corresponding Pi . If Pi is a 2 × 2 matrix, then from the Clairvoyance Lemma (Lemma 5.6) it is known that p QMA 1 − µi − µi(1 − µi) 1 Pi = p or . (5.52) − µi(1 − µi) µi

QMA QMA QMA If Pi has dimension 1, then Pi = 0, 1. If Pi is rank 0, 1 or 2, then EQMA the corresponding Pi is chosen to also be rank 0, 1 or 2 and of the same dimensions. From the deﬁnition of EQMA, we see that

0 0 P EQMA = 1 or (5.53) i 0 1 for rank 2 and rank 1 2 × 2 matrices respectively.

˜ Aside: the Hamiltonian HEQMA deﬁned here will generally be highly non-local.

32 To understand this, we note HEQMA is defined with respect to a QMA verification † ˜ circuit. However, this will not concern us since we only require that R HEQMAR † have a form and minimum eigenvalue close to R HQMAR to allow us to analyse the spectrum of HQMA more easily. We now analyse to subspaces of type 3 (as defined in Lemma 5.6).

Proving Energy Separation for YES and NO instances. We now consider how the {µi}i relate to the probability of witness rejection.

NO Instances: PTinit c Cd ⊗n The probability of a correctly initialised input |xi ∈ ker t=0 Πt = G ⊂ ( ) 0 L QMA 0 being rejected is hx| Πout|Gc |xi = hx | i Pi |Gc |x i ≤ 1. From the deﬁnition of 0 L QMA 0 QMA we have 1 − η ≤ hx | i Pi |x i ≤ 1 for a NO QMA instance. We note QMA L Pi |Gc = i µi. This implies 1 − η ≤ µi ≤ 1 for a Hamiltonian encoding the veriﬁcation of a NO instance. QMA QMA EQMA We now see that if Pi is one dimensional, Pi = Pi = 1 for NO QMA QMA instances. If Pi is a 2×2 matrix, then the Pi is as in expression 5.52. Thus EQMA QMA Pi − Pi = 0 or p EQMA QMA µi − 1 µi(1 − µi) (Pi − Pi )|KS ⊗HQ = p . (5.54) µi(1 − µi) 1 − µi

YES Instances: We now consider the similar argument for YES instances: By the definition of P QMA there must be at least one eigenvector |xi ∈ ker( t Πt) for which 0 ≤ L † QMA hx| i U Pi |Gc U |xi ≤ η. By the same reasoning we find there must be at QMA least one Pi = 0 or one 0 ≤ µi ≤ η. QMA The Pi are different from the NO instances as at least one witness must be accepted. Hence we know in the EQMA case that the ground state can receive no energy penalty, hence

1 0 0 0 1 0 0 0 P EQMA = or or or . (5.55) i 0 0 0 0 0 1 0 1

The ﬁrst two of these appear for witnesses that are accepted by the veriﬁer, and thus must occur for at least one i (as we are considering a YES instance). The third and fourth appear for rejected witnesses. We now consider the corresponding QMA cases (for convenience we will label µi = γi for witness that are accepted,

33 hence 0 ≤ γi ≤ η): p QMA 1 − γi − γi(1 − γi) 0 0 Pi = p or (5.56) − γi(1 − γi) γi 0 0 p 1 0 1 − µi − µi(1 − µi) or or p . (5.57) 0 1 − µi(1 − µi) µi

Here µi is only bounded between 1 − η ≤ µi ≤ 1. If we now consider accepting EQMA QMA witnesses we ﬁnd either Pi − Pi = 0 or p EQMA QMA γi γi(1 − γi) Pi − Pi = p . (5.58) γi(1 − γi) −γi

Energy Bound: We now consider the diﬀerence for the NO case. ˜ ∼ M EQMA QMA HEQMA|KS ⊗HQ − HQMA|KS ⊗HQ = |T i hT | ⊗ (Pi − Pi )|KS ⊗HQ i (5.59) ˜ (NO) (NO) EQMA QMA ||HEQMA|KS ⊗HQ − HQMA|KS ⊗HQ || = max ||Pi − Pi || (5.60) i 1/2 = max(1 − µi) (5.61) i ≤ η1/2. (5.62)

Using Lemma 5.11 we get:

˜ (NO) (NO) 1/2 |λ0 HEQMA|KS ⊗HQ − λ0 HQMA|KS ⊗HQ | ≤ η . (5.63) We now consider YES instances. To bound the minimum eigenvalue of Hamil- tonians encoding the veriﬁcation of YES instances we need only consider the minimum eigenvalue of the blocks corresponding to accepting witness(es). In these cases, γi ≤ η and µi ≥ 1 − η, hence

EQMA QMA 1/2 1/2 ||Pi − Pi || = max{γi , (1 − µi) } (5.64) ≤ η1/2. (5.65)

Thus

˜ (YES) (YES) 1/2 |λ0 HEQMA|KS ⊗HQ − λ0 HQMA |KS ⊗HQ | ≤ η . (5.66) Combining the bounds for both the YES and NO cases gives us:

˜ 1/2 |λ0 HEQMA|KS ⊗HQ − λ0 HQMA|KS ⊗HQ | ≤ η . (5.67)

34 Now consider the results in the statement of the lemma: the YES case is trivial ˜ (YES) (YES) as we know that λ0 HEQMA|KS ⊗HQ = 0 and HQMA |KS ⊗HQ ≥ 0, hence using the bound in equation (5.66),

(YES) 1/2 0 ≤ λ0 HQMA |KS ⊗HQ ≤ η . (5.68) To see the NO case, we note that the EQMA minimum eigenvalue occurs for a block of the form: ∆(T ) + |T i hT | . (5.69) Using the bound in equation (5.67), π π 1 − cos − η1/2 ≤λ H(NO)| ≤ 1 − cos . (5.70) 2T 0 QMA KS ⊗HQ 2T

Finally we note that η represents the probability of the veriﬁer being wrong. If we are interested in a QMA computation, then we can repeat the computation multiple times to get an exponentially better soundness and completeness boundaries [NC10][MW05]. We formalise this below. Corollary 5.13. Given a QMA instance, there exists a standard form Hamilto- nian, as described in Lemma 5.12, which has ground state energy in the bounds

(YES) −O(poly(n)) 0 ≤λ0 HQMA ≤ e (5.71) π π 1 − cos − e−O(poly(n)) ≤λ H(NO) ≤ 1 − cos . (5.72) 2T 0 QMA 2T Proof. We apply Lemma 5.12, where η is the probability of the QMA veriﬁer out- putting incorrectly: for YES instances η ≥ min|xi∈ker Hin hx| Πout |xi, while for NO instances 1 − η ≤ min|xi∈ker Hin hx| Πout |xi. If we are interested in a QMA computation, then we can repeat the computation a polynomial number of times to get an exponentially better soundness and completeness boundaries [NC10][MW05]. Using these ampliﬁcation methods, we amplify until η = O(e− poly(n)), and then apply Lemma 5.12, thus giving ground state energies exponentially close to the EQMA case.

6 Eigenvalue Scaling with Constant Rejection Prob- ability

In the previous section we used the fact that we can amplify in QMA in addition to perturbation theory to bound the minimum eigenvalues within an exponentially

35 Figure 3: Matrix shows the form of a standard form Hamiltonian H after conjugation by R. Each block has size T × T Note that Ai and Bij are diagonal † Tinit × Tinit sized matrices representing R HinR. The terms connect by solid blue lines are those result from the conjugation of Hout. The solid double-black line separates blocks which are inside and outside of the support of Hin. small region. However, it some cases there may be times we cannot amplify, in which case the η1/2 bound is not useful (since it may be the case η = 1/3). In this section we ﬁnd an upper bound that scale with T with constant η.

(YES) Theorem 6.1 (Constant Acceptance Upper Bound). Let HQMA be a standard form Hamiltonian with a standard form clock which encodes the veriﬁcation of a YES QMA instance. Let η = O(1) be the maximum probability of rejection as per the deﬁnition of QMA, then (YES) π η 0 ≤ λ0 HQMA ≤ η 1 − cos = O 2 . (6.1) 2(T − Tinit) + 1 T

(YES) Proof. From 5.12, subspaces of type 1 or 2 have minimum eigenvalue λ0 HQMA |KS ⊗HQ ≥ π 1 1 − cos O(n) = Ω n2 . Hence the only subspace where we can hope to get a low upper bound on the minimum eigenvalue is in the legal-only type 3 subspaces. For the remainder of this proof we only consider these type 3 subspaces.

36 We ﬁrst consider the matrix we are trying to bound the minimum eigenvalue of. Denote p 1 − µi − µi(1 − µi) P (µi) = p . (6.2) − µi(1 − µi) µi

Let KS ⊗ HQ be a type 3 subspace, then from the Clairvoyance Lemma (Lemma 5.6) with R = W (1C ⊗ (X ⊕ Y )Q), we have

† (YES) A :=R HQMA |KS ⊗HQ R (6.3) Tinit−1 M (T ) (T ) † X = (∆ ⊕ ∆ + 0T −1 ⊕ P (µi) ⊕ 0T −1) + R (|ti ht|) ⊗ (Πt) R. i t=0 (6.4) This has the structure seen in Figure 3. We now consider an inequality for the minimum eigenvalue using the Rayleigh quotient hν| A |νi λ0(A) ≤ min (6.5) |νi∈KS ⊗HQ hν|νi hν| A |νi ≤ min (6.6) |νi∈S⊆KS ⊗HQ hν|νi where S is some restricted subspace and going from the ﬁrst to second line is a consequence of the min-max theorem for matrix eigenvalues. We now consider the structure of A. We note that the blocks corresponding to each ∆(T ) which † PTinit−1 have support on Hin are coupled together by the R t=0 (|ti ht|) ⊗ (Πt) R term. Then the bottom-right blocks (i.e. those in the complement of the support (T ) of Hin) are disjoint from each other, but each is coupled to a single ∆ block in the top-left by a term P (µi). (T ) Consider the P (µi) with the smallest value of µi and the two ∆ blocks it couples together. Now restrict this subspace to everything except the ﬁrst Tinit rows and columns in the top-left block. This is now completely decoupled from rest of A. We label this matrix B0 and let the corresponding subspace it acts on be labelled S, such that dim S = T − Tinit. We see 1 B := (∆(T −Tinit) + |0i h0|) ⊕ ∆(T ) + 0 ⊕ P (µ) ⊕ 0 . (6.7) 2 T −Tinit−1 T −1 We now consider the inequality above and choose S to be the entire subspace Tinit−1 except the states {|ti}t=1 . Then from inequalities (6.5) and (6.6), we have λ0(A) ≤ λ0(B). We get the eigenvalue bound hν| B |νi λ0(A) ≤ min . (6.8) |νi∈S hν|νi

37 0 From now on denote T := T − Tinit. Further let the minimum eigenvector of ∆(T 0) + |0i h0| be |ui, where by Theorem 3.2 (viii) of [YC08] its components are given by

T 0 s T 0 X 2 sin(π/(4T 0 + 2)) X tπ |ui = u |ti = sin |t − 1i , (6.9) t sin(πT 0/(2T 0 + 1)) 2T 0 + 1 t=1 t=1 0 and the associated eigenvalue is γ0 = 1−cos(π/(2T + 1)). Furthermore, P (µ) has an eigenvector with zero eigenvalue √ µ |0i + p1 − µ |1i , (6.10) and ∆(T ) has an eigenvector with zero eigenvalue

T 1 X |wi = √ |ti . (6.11) T t=1 We then use the unnormalised vector √ ! 1 µ |ui |νi = √ uT√0 (6.12) T 1 − µ |wi as a trial ground state. Consider

2 hν| B |νi (µ/uT 0 )γ0 = 2 (6.13) hν|νi (µ/uT 0 ) + (1 − µ)T µγ0 = 2 . (6.14) µ + (1 − µ)T uT 0 We now note that 0 0 2 2T sin(π/(4T + 2)) 2 T π T |u 0 | = sin (6.15) T sin(πT 0/(2T 0 + 1)) 2T 0 + 1 π T 0π = 2T sin sin (6.16) 4T 0 + 2 2T 0 + 1 π π = 2T sin cos (6.17) 4T 0 + 2 4T 0 + 2 π = T sin (6.18) 2T 0 + 1

From the Clock Properties Assumptions (Assumptions 5.7) we know that Tinit = O(T 1/2), hence for T ≥ 2 we ﬁnd that

2 T |uT 0 | ≥ 1. (6.19)

38 Using this we get

hν| B |νi ≤ µγ . (6.20) hν|νi 0

We know that γ0 = 1 − cos(π/2(T − Tinit) + 1), hence, combining all of the above, we get

(YES) hν| H |νi π QMA ≤ µ 1 − cos . (6.21) hν|νi 2(T − Tinit) + 1

1/2 Again using Tinit = O(T ) we get

µ λ (H(YES)) = O . (6.22) 0 QMA T 2

Finally, we note µ ≤ η as η is the maximum probability of rejection of a correctly initialised witness, thus giving

η λ (H(YES)) = O . (6.23) 0 QMA T 2

7 Discussion and Outlook

The main aim of this work has been to understand the ground state eigenvalues and subspace as thoroughly as possible, as well as providing a toolbox for future work involving Feynman-Kitaev Hamiltonians and their extensions. We expect the constant rejection probability analysis to be useful for situations where our ability to provide ampliﬁcations is limited in some way. An example is in [BCW19] where we are allowed only very limited ampliﬁcation and we need to apply these bounds. We can then ask what else can we apply this analysis to:

Extensions to Unitary Labelled Graphs As mentioned previously, unitary labelled graphs representing branching computations have been shown to have promise gaps going as Ω(N −3) if the graph has N vertices [BCO17]. Other bounds are known, but are similarly fairly loose. Given some of the techniques in this paper (notably the Uncoupling Lemma (Lemma 4.1) allow us to decouple line graphs, it would be interesting to see if better bounds for ULGs can be found using these techniques or something similar.

39 Analysis of Quantum Walks on a Line The analysis quantum walks on a line in this paper is limited in that it only tells us that energy penalties at the end points are lower energy than elsewhere, and then given bounds on the energy of these end points. Intuitively, one would expect energy penalties closer to the centre of a computation to raise the energy more. It would be useful if we could rigorously understand how placing an energy penalty within a computation can change the energy — indeed such results would liked help us proving better bounds for ground state energies of unitary labelled Hamiltonians.

Fine-tuned Hamiltonian Energies A wider project can be seen in the context of designing Hamiltonians with specifically chosen energies and associated scalings for use in particular constructions. Examples include [BCW19] and [Bau+18], both of which rely on a construction where the energy of a negative energy Hamiltonian is traded-off against a positive energy Hamiltonian encoding a computation. There are two motivating points in this: extending the quantum walk analysis to bonus penalties to get a Hamiltonian which as an energy −f(T ), for runtime T , such that f(T ) only decays polynomially. The second point would be attempting to encode a computation in a Hamiltonian with negative ground state energies. At the moment it is not known how to do this due to difficulties initialising the encoded circuit/quantum Turing Machine. That is, the bonus provided by reaching an accepting state is usually sufficiently large to make it favourable to pick up energy penalties in the initialisation steps.

8 Acknowledgements

The author would like to thank Toby Cubitt for support and useful discussions particularly regarding Section 5, and Johannes Bausch for discussions regarding the constant acceptance probability lemma. The author is supported by the EPSRC Centre for Doctoral Training in Delivering Quantum Technologies.

References

[AA11] S. Aarsonson and A. Arkipov “The computational complexity of linear optics”. In: Proceedings of the 43rd annual ACM symposium on Theory of Computing - STOC 11 (2011).

[ADH05] Leonard M. Adleman, Jonathan DeMarrais, and Ming-Deh A. Huang. “Quantum computability”. In: SIAM Journal on Computing (2005).

[AAV13] “Guest Column”. In: ACM SIGACT News 44.2 (Mar. 2013), p. 47.

40 [Aha+08] Dorit Aharonov, Wim Van Dam, Julia Kempe, Zeph Landau, Seth Lloyd, and Oded Regev. “Adiabatic Quantum Computation Is Equiv- alent to Standard Quantum Computation”. In: SIAM Review 50.4 (2008), pp. 755-787.

[Aha+09] Dorit Aharonov, Daniel Gottesman, Sandy Irani, and Julia Kempe. “The Power of Quantum Systems on a Line”. In: Communications in Mathematical Physics 287.1 (Apr. 2009), pp. 41-65. arXiv: 0705.4077 [quant-ph].

[Amb14] Andris Ambainis. “On Physical Problems that are Slightly More Dif- ﬁcult than QMA”. In: 2014 IEEE 29th Conference on Computational Complexity (CCC) (2014).

[Bau+18] J. Bausch, T. S. Cubitt, A. Lucia, and D. Perez-Garcia. “Undecidability of the Spectral Gap in One Dimension”. In: (2018). arXiv: 1810.01858 [quant-ph].

[BCW19] J. Bausch, T. S. Cubitt, and J. D. Watson. “Uncomputability of Phase Diagrams”. In: To be published (2019).

[BC18] Johannes Bausch and Elizabeth Crosson. “Analysis and limitations of modi ed circuit-to-Hamiltonian constructions”. In: Quantum 2 (Sept. 2018), p. 94. arXiv: 1609.08571.

[BCO17] Johannes Bausch, Toby Cubitt, and Maris Ozols. “The Complexity of Translationally-Invariant Spin Chains with Low Local Dimension”. In: Annales Henri Poincare (2017), p. 52. arXiv: 1605.01718.

[BV97] E Bernstein and U Vazirani. “Quantum complexity theory”. In: SIAM Journal on Computing 26.5 (1997), pp. 1411-1473.

[Boh+19] Thomas C. Bohdanowicz, Elizabeth Crosson, Chinmay Nirkhe, and Henry Yuen. “Good approximate quantum LDPC codes from space- time circuit Hamiltonians”. In: Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing - STOC 2019 (2019).

[NT14] Nikolas P Breuckmann and Barbara M Terhal. “Space-time circuit-to- Hamiltonian construction and its applications”. In: Journal of Physics A: Mathematical and Theoretical 47.19 (2014), p. 195304.

41 [BFS11] Brielin Brown, Steven T. Flammia, and Norbert Schuch. “Computa- tional Difficulty of Computing the Density of States”. In: Physical Re- view Letters 107.4 (2011). [CLN18] Libor Caha, Zeph Landau, and Daniel Nagaj. “Clocks in Feynman’s computer and Kitaev’s local Hamiltonian: Bias, gaps, idling, and pulse tuning”. In: Physical Review A 97.6 (June 2018), p. 062306. arXiv: 1712.07395. [CB17] Elizabeth Crosson and John Bowen. “Quantum ground state isoperi- metric inequalities for the energy spectrum of local Hamiltonians’. In: arXiv e-prints, arXiv:1703.10133 (Mar. 2017), arXiv:1703.10133. arXiv: 1703.10133 [quant-ph]. [CPW15] T. S. Cubitt, D. Perez-Garcia, and M. M. Wolf. “Undecidability of the spectral gap”. In: (2015). arXiv: 1502.04573 [quant-ph]. [CM13] Toby Cubitt and Ashley Montanaro. “Complexity classification of local Hamiltonian problems”. In: arXiv e-prints, arXiv:1311.3161 (Nov. 2013), arXiv:1311.3161. arXiv: 1311.3161 [quant-ph]. [GS15] Sevag Gharibian and Jamie Sikora.“Ground State Connectivity of Local Hamiltonians”. In: Automata, Languages, and Programming Lecture Notes in Computer Science (2015), pp. 617-628. [GY16] S.Gharibian and J. Yirka. “The complexity of simulating local measure- ments on quantum systems”. In: (2016). arXiv: 1606.05626 [quant-ph]. [GC18] Carlos E. González-Guillénand Toby S. Cubitt. “History-state Hamil- tonians are critical”. In: arXiv e-prints, arXiv:1810.06528 (Oct. 2018), arXiv:1810.06528. arXiv: 1810.06528 [quant-ph]. [GI09] Daniel Gottesman and Sandy Irani. “The quantum and classical complexity of translationally invariant tiling and Hamiltonian problems”. In: Foundations of Computer Science, 2009. FOCS’09. 50th Annual IEEE Symposium on. IEEE. 2009, pp. 95-104. [JGL10] Stephen P. Jordan, David Gosset, and Peter J. Love. “Quantum-Merlin- Arthur-complete problems for stoquastic Hamiltonians and Markov matrices”. In: Physical Review A 81.3 (2010). [KKR04] Julia Kempe, Alexei Kitaev, and Oded Regev. “The Complexity of the Local Hamiltonian Problem”. In: FSTTCS 2004: Foundations of Soft- ware Technology and Theoretical Computer Science Lecture Notes in Computer Science (2004), pp. 372-383.

42 [KSV02] A. Yu. Kitaev, A. Shen, and M. N. Vyalyi. Classical and quantum computation. American Mathematical Society, 2002.

[MW05] C. Marriott and J. Watrous. “Quantum Arthur-Merlin games”. In: Pro- ceedings. 19th IEEE Annual Conference on Computational Complexity, 2004. (2005).

[NC10] Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Information. Cambridge: Cambridge University Press, 2010, p. 676.

[OT08] R. Oliveira and B. M. Terhal. “The complexity of quantum spin systems on a two dimensional square lattice”. In: Quantum Information and Computation 8.10 (Nov. 2008), pp. 0900-0924.

[UHB17] Na¨ıriUsher, Matty J. Hoban, and Dan E. Browne. “Nonunitary quantum computation in the ground space of local Hamiltonians”. In: Phys- ical Review A 96.3 (Dec. 2017).

[Yue05] Wen-Chyuan Yueh. “Eigenvalues of several tridiagonal matrices”. In: Applied Mathematics E-notes. 2005, pp. 5-66.

[YC08] Wen-Chyuan Yueh and Sui Sun Cheng. “Explicit Eigenvalues And In- verses Of Tridiagonal Toeplitz Matrices With Four Perturbed Corners”. In: The ANZIAM Journal 49.03 (2008), p. 361.